Inside 360’s Image Service: How Davinci and Picasso Power Fast Uploads and Downloads

Overview

Many companies build shared infrastructure services such as push, cache, MQ, storage, and image hosting to improve service quality and development speed. 360’s image service, called HULK, supports most business lines with features like cropping, compression, filters, pHash calculation, face detection, format conversion, and GIF first‑frame extraction, handling over 150 billion page views per day.

Upload Module (Davinci)

The upload process consists of the following steps:

SDK upload : Clients use an SDK or direct HTTP POST to the Davinci upload API, which is load‑balanced by Nginx on port 80.

Nginx 80 forwarding : Requests are forwarded via upstream to backend Nginx on port 8360.

Queueing : The backend service enqueues the upload task to Gearman, returning a task ID for result queries.

Image processing and storage : Gearman workers asynchronously compress, crop, perform face detection, etc., then store the image and its metadata in Cassandra; the processing result is cached in Redis.

Result retrieval : Clients poll the task ID to obtain the final image URL and metadata. Synchronous upload and callback notifications are also supported.

Download Module (Picasso)

The download flow works as follows:

Image URL request : The client requests an image URL, which is resolved via CNAME to a CDN node. If the CDN has the image cached, it returns the data immediately.

CDN miss : The request is sent back to the backend Nginx on port 80.

Varnish front‑cache : Varnish checks its cache; a hit returns the image directly.

Backend handling : On a miss, Nginx on port 8360 forwards the request to PHP‑FastCGI, which reads the image from storage.

Image processing : The Nginx filter module (built with GraphicsMagick) applies transformations specified in the URL parameters, such as cropping, filters, black‑white conversion, face cropping, or extracting the first frame of a GIF.

Response : The processed image is returned to the client and cached at the CDN edge.

Supporting Infrastructure

The service relies on LVS for load balancing, Redis for caching task results, Cassandra for persistent storage, CDN for edge delivery, QCMD for command execution, and the Wonder monitoring system, forming a stable and maintainable backend ecosystem.